I wonder how much of that actually survives token filtering during training